Query will complete, but in aprox 3,5 to 4 min, you will need up to 14 GB of memory. (4 Core, Win10, 32GB RAM).
You can see below, memory usage in MB, throughout the query, sampled at 15 sec interval.
duckdb 321.01 -> Start Query
duckdb 6302.12
duckdb 13918.04
duckdb 10963.74
duckdb 8586.76
duckdb 7613.86
duckdb 6749.53
duckdb 5990.96
duckdb 5293.35
duckdb 4205.53
duckdb 3153.59
duckdb 1482.86
duckdb 386.29 -> End Query
Indeed, 14GB seems really high for a 400MB Parquet file, that's a 35x multiple on the base file size.
Of course, the data is compressed on disk, but even the uncompressed data isn't that large so I believe indeed that quite a lot of optimisations are still possible.
Newer DuckDbs are able to handle out of core operations better. But in general just because data fits in memory doesn’t mean the operation will — and as I said 8GB is very limited memory so it will entail spilling to disk.
Parquet files are compressed, and many analytic operations require more memory than the on disk size. When you don’t have enough ram DuckDb has to switch to out of core mode which is slower. (It’s the classic performance trade off)
8gb of ram is not enough usually to expect performance from analytic operations - I usually have minimum 16. My current instance is remote which has 256gb ram. I never run out of ram and DuckDb never fails and runs super fast.
``` COPY ( SELECT DISTINCT ON (b, c) * FROM READ_PARQUET('input.parquet') ORDER BY a DESC ) TO 'output.parquet' ( FORMAT PARQUET, COMPRESSION 'ZSTD' ) ; ```
Where the input file has 25M rows (500Mb in Parquet format) containing 4 columns, a and b are BIGINTs and c and d are VARCHARs.
On a Mac Book Pro M1 with 8GB of RAM (16x the original file size), the query will not finish.
This is a query that could very easily be optimised to take little amounts of space (hash the DISTINCT ON key, and replace in-place the already seen values if the value of "a" is larger than the one that already exists.)